16×NVIDIA H200 deployment of GLM-5.2

For an anonymised Shenzhen enterprise offline environment, the JSLE team delivered a two-node GPU cluster, GLM-5.2-FP8 inference, a unified API route and a governed agent workspace, then completed functional verification.

Successful delivery case · Deployment, functional verification and acceptance complete

GPU scale2 inference nodes · 16 H200 GPUs

Target modelGLM-5.2-FP8

Inference stackRay + vLLM across two nodes

EnvironmentCustomer network · Offline implementation

Project facts

Background and boundaries

The customer wanted controlled inference in its own data centre, with model weights, invocation paths and runtime logs kept in the enterprise network. Client name, domains, people, server addresses and security settings are redacted; this page presents only the anonymised architecture, delivery method and test conclusions.

Functional verification

All six basic functional tests passed. Four concurrent SSE streams were verified and independent sessions remained isolated. Performance benchmarks, long-context behaviour and higher concurrency remain subject to agreed test conditions and formal acceptance evidence.

Next step

For a multi-node planning discussion, we can provide capacity sizing, implementation and acceptance planning.

Phone: +86 139 2521 1225 Email: tianjun@jsle.cn